Aleksey Makarov says:
====================
net: thunderx: Misc fixes
Miscellaneous fixes for the ThunderX VNIC driver
All the patches can be applied individually.
It's ok to drop some if the maintainer feels uncomfortable
with applying for 4.2.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Cortina phy does not have kernel driver and we don't attach
device with phy layer for intefaces like XFI, XLAUI etc,
Hence check for interface type before calling disconnect.
Signed-off-by: Thanneeru Srinivasulu <tsrinivasulu@caviumnetworks.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes a crash when changing rss with multiple traffic flows.
While interface teardown, disable tx queues after all NAPI threads
are done. If done otherwise tx queues might be woken up inside NAPI
if any CQE_TX are processed.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a txq (SQ) remains in stopped state after this timeout its
considered as stuck and interface is reinited.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously TXQ is wakedup whenever napi is executed
and irrespective of if any CQE_TX are processed or not.
Added 'txq_stop' and 'txq_wake' counters to aid in debugging
if there are any future issues.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Suppressing standard alloc_pages() warnings. Some kernel configs limit
alloc size and the network driver may fail. Do not drop a kernel
warning in this case, instead just drop a oneliner that the network
driver could not be loaded since the buffer could not be allocated.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixing TSO packages not being counted.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix for memory leak when changing queue/channel count via ethtool
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With earlier configured value sufficient number of CQEs are not
being reserved for transmitted packets. Hence under heavy incoming
traffic load, receive notifications will take away most of the CQ
thus transmit notifications will be lost resulting in tx skbs not
being freed.
Finally SQ will be full and it will be stopped, watchdog timer
will kick in. After this fix receive notifications will not take
morethan half of CQ reserving the rest for transmit notifications.
Also changed CQ & SQ sizes from 16k to 4k.
This is also due to the receive notifications taking first half of
CQ under heavy load and time taken by NAPI to clear transmit notifications
will increase with higher queue sizes. Again results in SQ being stopped.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixed 'tso_hdrs' memory not being freed properly.
Also fixed SQ skbuff maintenance issues.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Switching back to LDD transactions from LDWB.
While transmitting packets out with LDWB transactions
data integrity issues are seen very frequently.
hence switching back to LDD.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: Robert Richter <rrichter@cavium.com>
Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch is the IPv6 equivalent of commit
6c8b4e3ff8 ("arp: flush arp cache on IFF_NOARP change")
Without it, we keep buggy neighbours in the cache, with destination
MAC address equal to our own MAC address.
Tested:
tcpdump -i eth0 -s 0 ip6 -n -e &
ip link set dev eth0 arp off
ping6 remote // sends buggy frames
ip link set dev eth0 arp on
ping6 remote // should work once kernel is patched
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Mario Fanelli <mariofanelli@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Murali Karicheri says:
====================
net: netcp: bug fixes for dynamic module support
This series fixes few bugs to allow keystone netcp modules to be
dynamically loaded and removed. Currently it allows following
sequence multiple times
insmod cpsw_ale.ko
insmod davinci_mdio.ko
insmod keystone_netcp.ko
insmod keystone_netcp_ethss.ko
ifup eth0
ifup eth1
ping <hosts on eth0>
ping <hosts on eth1>
ifdown eth1
ifdown eth0
rmmod keystone_netcp_ethss.ko
rmmod keystone_netcp.ko
rmmod davinci_mdio.ko
rmmod cpsw_ale.ko
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch clean up error handle code to use goto label properly. In some
cases, the code unnecessarily use goto instead of just returning the error
code. Code also make explicit calls to devm_* APIs on error which is
not necessary. In the gbe_remove() also it makes similar calls which is
also unnecessary.
Also fix few checkpatch warnings
Signed-off-by: Murali Karicheri <m-karicheri2@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The code seems to assume a null is returned when the list is empty
from first_sec_slave() to break the loop which is incorrect. Fix the
code by using list_empty().
Signed-off-by: Murali Karicheri <m-karicheri2@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently if user do rmmod keystone_netcp.ko following warning is
seen :-
[ 59.035891] ------------[ cut here ]------------
[ 59.040535] WARNING: CPU: 2 PID: 1619 at drivers/net/ethernet/ti/
netcp_core.c:2127 netcp_remove)
This is because the interface list is not cleaned up in netcp_remove.
This patch fixes this. Also fix some checkpatch related warnings.
Signed-off-by: Murali Karicheri <m-karicheri2@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since mdb states were introduced when deleting an entry the state was
left as it was set in the delete request from the user which leads to
the following output when doing a monitor (for example):
$ bridge mdb add dev br0 port eth3 grp 239.0.0.1 permanent
(monitor) dev br0 port eth3 grp 239.0.0.1 permanent
$ bridge mdb del dev br0 port eth3 grp 239.0.0.1 permanent
(monitor) dev br0 port eth3 grp 239.0.0.1 temp
^^^
Note the "temp" state in the delete notification which is wrong since
the entry was permanent, the state in a delete is always reported as
"temp" regardless of the real state of the entry.
After this patch:
$ bridge mdb add dev br0 port eth3 grp 239.0.0.1 permanent
(monitor) dev br0 port eth3 grp 239.0.0.1 permanent
$ bridge mdb del dev br0 port eth3 grp 239.0.0.1 permanent
(monitor) dev br0 port eth3 grp 239.0.0.1 permanent
There's one important note to make here that the state is actually not
matched when doing a delete, so one can delete a permanent entry by
stating "temp" in the end of the command, I've chosen this fix in order
not to break user-space tools which rely on this (incorrect) behaviour.
So to give an example after this patch and using the wrong state:
$ bridge mdb add dev br0 port eth3 grp 239.0.0.1 permanent
(monitor) dev br0 port eth3 grp 239.0.0.1 permanent
$ bridge mdb del dev br0 port eth3 grp 239.0.0.1 temp
(monitor) dev br0 port eth3 grp 239.0.0.1 permanent
Note the state of the entry that got deleted is correct in the
notification.
Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Fixes: ccb1c31a7a ("bridge: add flags to distinguish permanent mdb entires")
Signed-off-by: David S. Miller <davem@davemloft.net>
When fast leave is configured on a bridge port and an IGMP leave is
received for a group, the group is not deleted immediately if there is
a router detected or if multicast querier is configured.
Ideally the group should be deleted immediately when fast leave is
configured.
Signed-off-by: Satish Ashok <sashok@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are several devices that can receive vlan tagged packets with
CHECKSUM_PARTIAL like tap, possibly veth and xennet.
When (multiple) vlan tagged packets with CHECKSUM_PARTIAL are forwarded
by bridge to a device with the IP_CSUM feature, they end up with checksum
error because before entering bridge, the network header is set to
ETH_HLEN (not including vlan header length) in __netif_receive_skb_core(),
get_rps_cpu(), or drivers' rx functions, and nobody fixes the pointer later.
Since the network header is exepected to be ETH_HLEN in flow-dissection
and hash-calculation in RPS in rx path, and since the header pointer fix
is needed only in tx path, set the appropriate network header on forwarding
packets.
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Signed-off-by: David S. Miller <davem@davemloft.net>
tpacket_fill_skb() can return a negative value (-errno) which
is stored in tp_len variable. In that case the following
condition will be (but shouldn't be) true:
tp_len > dev->mtu + dev->hard_header_len
as dev->mtu and dev->hard_header_len are both unsigned.
That may lead to just returning an incorrect EMSGSIZE errno
to the user.
Fixes: 52f1454f62 ("packet: allow to transmit +4 byte in TX_RING slot for VLAN case")
Signed-off-by: Alexander Drozdov <al.drozdov@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
When arp is off on a device, and ioctl(SIOCGARP) is queried,
a buggy answer is given with MAC address of the device, instead
of the mac address of the destination/gateway.
We filter out NUD_NOARP neighbours for /proc/net/arp,
we must do the same for SIOCGARP ioctl.
Tested:
lpaa23:~# ./arp 10.246.7.190
MAC=00:01:e8:22:cb:1d // correct answer
lpaa23:~# ip link set dev eth0 arp off
lpaa23:~# cat /proc/net/arp # check arp table is now 'empty'
IP address HW type Flags HW address Mask Device
lpaa23:~# ./arp 10.246.7.190
MAC=00:1a:11:c3:0d:7f // buggy answer before patch (this is eth0 mac)
After patch :
lpaa23:~# ip link set dev eth0 arp off
lpaa23:~# ./arp 10.246.7.190
ioctl(SIOCGARP) failed: No such device or address
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Vytautas Valancius <valas@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This notification causes the FIB to be updated, which is not needed
because the address already exists, and more importantly it may undo
intentional changes that were made to the FIB after the address was
originally added. (As a point of comparison, when an address becomes
deprecated because its preferred lifetime expired, a notification on
this chain is not generated.)
The motivation for this commit is fixing an incompatibility between
DHCP clients which set and update the address lifetime according to
the lease, and a commercial VPN client which replaces kernel routes
in a way that outbound traffic is sent only through the tunnel (and
disconnects if any further route changes are detected via netlink).
Signed-off-by: David Ward <david.ward@ll.mit.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
These should be handled only by the respective STP which is in control.
They become problematic for devices with limited resources with many
ports because the hold_timer is per port and fires each second and the
hello timer fires each 2 seconds even though it's global. While in
user-space STP mode these timers are completely unnecessary so it's better
to keep them off.
Also ensure that when the bridge is up these timers are started only when
running with kernel STP.
Signed-off-by: Satish Ashok <sashok@cumulusnetworks.com>
Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When binding a PF_PACKET socket, the use count of the bound interface is
always increased with dev_hold in dev_get_by_{index,name}. However,
when rebound with the same protocol and device as in the previous bind
the use count of the interface was not decreased. Ultimately, this
caused the deletion of the interface to fail with the following message:
unregister_netdevice: waiting for dummy0 to become free. Usage count = 1
This patch moves the dev_put out of the conditional part that was only
executed when either the protocol or device changed on a bind.
Fixes: 902fefb82e ('packet: improve socket create/bind latency in some cases')
Signed-off-by: Lars Westerhoff <lars.westerhoff@newtec.eu>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Network header is set with offset ETH_HLEN but it is not true for VLAN
(multiple-)tagged and results in checksum issues in lower devices.
v2: leave skb->protocol untouched (thx Vlad), comment added
v3: moved after skb_probe_transport_header() call (thx Toshiaki)
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It was reported that update_suffix was taking a long time on systems where
a large number of leaves were attached to a single node. As it turns out
fib_table_flush was calling update_suffix for each leaf that didn't have all
of the aliases stripped from it. As a result, on this large node removing
one leaf would result in us calling update_suffix for every other leaf on
the node.
The fix is to just remove the calls to leaf_pull_suffix since they are
redundant as we already have a call in resize that will go through and
update the suffix length for the node before we exit out of
fib_table_flush or fib_table_flush_external.
Reported-by: David Ahern <dsa@cumulusnetworks.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Tested-by: David Ahern <dsa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If an architecture defines readl/writel using CPP macros, we
get the following kinds of build failure:
> > > drivers/net/ethernet/cadence/macb.c:164:1: error: macro "writel"
> > > passed 3 arguments, but takes just 2
> macb_or_gem_writel(bp, SA1B, bottom);
> ^
Rename the methods so that this doesn't happen.
Signed-off-by: David S. Miller <davem@davemloft.net>
When a switch is attached to the mdio bus, the mdio bus can be used
while the interface is not open. If the IPG clock is not enabled, MDIO
reads/writes will simply time out.
Add support for runtime PM to control this clock. Enable/disable this
clock using runtime PM, with open()/close() and mdio read()/write()
function triggering runtime PM operations. Since PM is optional, the
IPG clock is enabled at probe and is no longer modified by
fec_enet_clk_enable(), thus if PM is not enabled in the kernel, it is
guaranteed the clock is running when MDIO operations are performed.
Signed-off-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Cc: tyler.baker@linaro.org
Cc: fabio.estevam@freescale.com
Cc: shawn.guo@linaro.org
Tested-by: Fabio Estevam <fabio.estevam@freescale.com>
Tested-by: Tyler Baker <tyler.baker@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch asserts SGMII RTRESET, i.e. resetting the SGMII Tx/Rx
logic, during network interface shutdown to avoid having the
hardware wedge when shutting down with high incoming traffic rates.
This is cleared (brought out of RTRESET) when the interface is
brought back up.
Signed-off-by: WingMan Kwok <w-kwok2@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Andy Shevchenko says:
====================
net/macb: fix for AVR32 and clean up
It seems no one had tested recently the driver on AVR32 platforms such as
ATNGW100. This series bring it back to work.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch coverts struct description to the kernel doc format. There is no
functional change.
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
macb_count_tx_descriptors() repeats the generic macro DIV_ROUND_UP(). The patch
does a replacement.
There is no functional change.
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes the following warnings:
drivers/net/ethernet/cadence/macb.c: In function ‘macb_handle_link_change’:
drivers/net/ethernet/cadence/macb.c:266: warning: comparison between signed and unsigned
drivers/net/ethernet/cadence/macb.c:267: warning: comparison between signed and unsigned
drivers/net/ethernet/cadence/macb.c:291: warning: comparison between signed and unsigned
drivers/net/ethernet/cadence/macb.c: In function ‘gem_update_stats’:
drivers/net/ethernet/cadence/macb.c:1908: warning: comparison between signed and unsigned
drivers/net/ethernet/cadence/macb.c: In function ‘gem_get_ethtool_strings’:
drivers/net/ethernet/cadence/macb.c:1988: warning: comparison between signed and unsigned
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To avoid messages like
macb macb.0 (unnamed net_device) (uninitialized): Cadence caps 0x00000000
macb macb.0 (unnamed net_device) (uninitialized): invalid hw address, using random
let's use dev_*() macros.
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit 98b5a0f4a2 introduces jumbo frame support, but also it assumes
that macb_config present which is not always true.
The configuration without macb_config fails to boot.
Unable to handle kernel NULL pointer dereference at virtual address 00000010
ptbr = 90350000 pgd = 00000000
Oops: Kernel access of bad area, sig: 11 [#1]
FRAME_POINTER chip: 0x01f:0x1e82 rev 2
Modules linked in:
CPU: 0 PID: 1 Comm: swapper Not tainted 4.2.0-rc3-next-20150723+ #13
task: 91c26000 ti: 91c28000 task.ti: 91c28000
PC is at macb_probe+0x140/0x61c
Fixes: 98b5a0f4a2 (net: macb: Add support for jumbo frames)
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit a50dad355a (net: macb: Add big endian CPU support) converted I/O
accessors to readl_relaxed() and writel_relaxed() and consequentially broke
MACB driver on AVR32 platforms such as ATNGW100.
This patch improves I/O access by checking endiannes first and use the
corresponding methods.
Fixes: a50dad355a (net: macb: Add big endian CPU support)
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, tcp_recvmsg enters a busy loop in sk_wait_data if called
with flags = MSG_WAITALL | MSG_PEEK.
sk_wait_data waits for sk_receive_queue not empty, but in this case,
the receive queue is not empty, but does not contain any skb that we
can use.
Add a "last skb seen on receive queue" argument to sk_wait_data, so
that it sleeps until the receive queue has new skbs.
Link: https://bugzilla.kernel.org/show_bug.cgi?id=99461
Link: https://sourceware.org/bugzilla/show_bug.cgi?id=18493
Link: https://bugzilla.redhat.com/show_bug.cgi?id=1205258
Reported-by: Enrico Scholz <rh-bugzilla@ensc.de>
Reported-by: Dan Searle <dan@censornet.com>
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hayes Wang says:
====================
r8152: issues fix
v2:
Replace patch #2 with "r8152: fix wakeup settings".
v1:
These patches are used to fix issues.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Adjust napi_disable() and napi_enable() to avoid r8152_poll() start
working before rx ready. Otherwise, it may have race condition for
rx_agg.
Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Avoid the driver to enable WOL if the device doesn't support it.
Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
- Disable U1/U2 during initialization.
- Disable lpm when linking is on, and enable it when linking is off.
- Disable U1/U2 when enabling runtime suspend.
It is possible to let hw stop working, if the U1/U2 request occurs
during some situations. The patch is used to avoid it.
Signed-off-by: Hayes Wang <hayeswang@realtek.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Johan Hedberg says:
====================
pull request: bluetooth 2015-07-23
Here's another one-patch pull request for 4.2 which targets a potential
NULL pointer dereference in the LE Security Manager code that can be
triggered by using older user space tools. The issue has been there
since 4.0 so there's the appropriate "Cc: stable" in place.
Let me know if there are any issues pulling. Thanks.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
This function frees resources and cancels delayed work item that
have been initialized in fec_ptp_init().
Use this to do proper error handling if something goes wrong in
probe function after fec_ptp_init has been called.
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Acked-by: Fugang Duan <B38611@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
So it gets freed when the device is going away.
This fixes a DMA memory leak on driver probe() fail and driver
remove().
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes: a3138df9 ("[NIU]: Add Sun Neptune ethernet driver.")
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian Westphal says:
====================
inet: ip defrag bug fixes
Johan Schuijt and Frank Schreuder reported crash and softlockup after the
inet workqueue eviction change:
general protection fault: 0000 [#1] SMP
CPU: 0 PID: 4 Comm: kworker/0:0 Not tainted 3.18.18-transip-1.5 #1
Workqueue: events inet_frag_worker
task: ffff880224935130 ti: ffff880224938000 task.ti: ffff880224938000
RIP: 0010:[<ffffffff8149288c>] [<ffffffff8149288c>] inet_evict_bucket+0xfc/0x160
RSP: 0018:ffff88022493bd58 EFLAGS: 00010286
RAX: ffff88021f4f3e80 RBX: dead000000100100 RCX: 000000000000006b
RDX: 000000000000006c RSI: ffff88021f4f3e80 RDI: dead0000001000a8
RBP: 0000000000000002 R08: ffff880222273900 R09: ffff880036e49200
R10: ffff8800c6e86500 R11: ffff880036f45500 R12: ffffffff81a87100
R13: ffff88022493bd70 R14: 0000000000000000 R15: ffff8800c9b26280
[..]
Call Trace:
[<ffffffff814929e0>] ? inet_frag_worker+0x60/0x210
[<ffffffff8107e3a2>] ? process_one_work+0x142/0x3b0
[<ffffffff8107eb94>] ? worker_thread+0x114/0x440
[..]
A second issue results in softlockup since the evictor may restart the
eviction loop for a (potentially) unlimited number of times while local
softirqs are disabled.
Frank reports that test system remained stable for 14 hours of testing
(before, crash occured within half an hour in their setup).
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
We can simply remove the INET_FRAG_EVICTED flag to avoid all the flags
race conditions with the evictor and use a participation test for the
evictor list, when we're at that point (after inet_frag_kill) in the
timer there're 2 possible cases:
1. The evictor added the entry to its evictor list while the timer was
waiting for the chainlock
or
2. The timer unchained the entry and the evictor won't see it
In both cases we should be able to see list_evictor correctly due
to the sync on the chainlock.
Joint work with Florian Westphal.
Tested-by: Frank Schreuder <fschreuder@transip.nl>
Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Frank reports 'NMI watchdog: BUG: soft lockup' errors when
load is high. Instead of (potentially) unbounded restarts of the
eviction process, just skip to the next entry.
One caveat is that, when a netns is exiting, a timer may still be running
by the time inet_evict_bucket returns.
We use the frag memory accounting to wait for outstanding timers,
so that when we free the percpu counter we can be sure no running
timer will trip over it.
Reported-and-tested-by: Frank Schreuder <fschreuder@transip.nl>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller
Thanneeru Srinivasulu
Sunil Goutham
Eric Dumazet
Karicheri, Muralidharan
Daniel Borkmann
Nikolay Aleksandrov
Satish Ashok
Toshiaki Makita
Alexander Drozdov
David Ward
Nikolay Aleksandrov
Lars Westerhoff
Ivan Vecera
Alexander Duyck
Andrew Lunn
WingMan Kwok
Andy Shevchenko
Sabrina Dubroca
hayeswang
Lucas Stach
Jiri Pirko
Florian Westphal